Adaptive Negative Textual Space (ANTS)

• Adaptive Negative Textual Space (ANTS) is a framework for OOD detection that constructs test-time negative textual cues using mined images.
• It leverages MLLMs to generate two complementary negative texts—expressive sentences for far-OOD and labels for near-OOD detection.
• ANTS combines CLIP similarity scores with an adaptive weighted rule to balance negative texts and reduce false negatives across varying OOD regimes.

Searching arXiv for the primary ANTS paper and closely related work on adaptive negative labels/textual semantics for OOD detection. Searching arXiv for "Adaptive Negative Textual Space OOD detection MLLM" and related negative-label OOD methods. Adaptive Negative Textual Space (ANTS) is a framework for out-of-distribution (OOD) detection with vision-LLMs that constructs an adaptive negative textual space at inference time by leveraging the understanding and reasoning capabilities of multimodal LLMs (MLLMs) (Wenjie et al., 4 Sep 2025). The method is motivated by two limitations of existing negative-label approaches: they often lack an understanding of OOD images, which makes it difficult to construct an accurate negative space, and false negative labels significantly degrade near-OOD performance (Wenjie et al., 4 Sep 2025). ANTS addresses these issues by mining likely OOD images from the test set, generating two complementary forms of negative text from those images, and combining them with an adaptive weighted score that is intended to handle both near-OOD and far-OOD settings without task-specific prior knowledge (Wenjie et al., 4 Sep 2025).

1. Problem setting and motivation

Negative labels have been shown to enhance OOD detection, but static or poorly targeted negative labels can fail to cover varied OOD semantics and can introduce false negatives when OOD samples resemble a subset of in-distribution (ID) classes (Wenjie et al., 4 Sep 2025). In ANTS, the central object is the “adaptive negative textual space”: a negative text space shaped from test-time evidence rather than fixed in advance. The framework operates with off-the-shelf pre-trained CLIP and MLLMs, and it is explicitly described as training-free and zero-shot, with all negative textual space construction occurring at test time (Wenjie et al., 4 Sep 2025).

The framework distinguishes two regimes. In far-OOD settings, OOD samples are very dissimilar to the ID distribution, so rich descriptive negative text is advantageous. In near-OOD settings, OOD samples resemble part of the ID label set, so generic or overly broad negative descriptions can create false negatives. ANTS therefore builds one negative textual space for expressive characterization of OOD semantics and another that is tailored to the visually similar ID subset (Wenjie et al., 4 Sep 2025). This suggests that the method treats OOD detection as a problem of matching the granularity of negative text to the geometry of the encountered shift.

2. Negative image mining at test time

ANTS begins by selecting images likely to be OOD from the test set and treating them as negative images. A baseline OOD detector, such as NegLabel, assigns each test image $x$ a score $S_{nl}(x)$, and images with low similarity to ID labels or high similarity to negative labels are selected as candidates (Wenjie et al., 4 Sep 2025). The negative-image set is defined as

$$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$

To avoid manual thresholding, ANTS uses an adaptive filtering scheme. Rather than fixing $\gamma$ directly, it selects the bottom percentile $\eta$ of test samples with the lowest $S_{nl}$, which induces a data-driven dynamic threshold $\gamma^*$ (Wenjie et al., 4 Sep 2025). The detailed summary gives “e.g., 50%” as an example of the bottom percentile used for this adaptive filtering (Wenjie et al., 4 Sep 2025).

This stage is significant because the negative textual space is not synthesized from an external ontology or a fixed vocabulary alone. It is grounded in mined negative images drawn from the current test distribution. A common misconception is that ANTS is merely a larger pool of static negative labels; in fact, the framework first infers likely negatives from the test stream and only then asks the MLLM to construct negative text around those images (Wenjie et al., 4 Sep 2025).

3. Two complementary negative textual spaces

From the mined negative images, ANTS constructs two types of negative text: Expressive Negative Sentences (ENS) and Visually Similar Negative Labels (VSNL) (Wenjie et al., 4 Sep 2025). ENS targets far-OOD discrimination, whereas VSNL is designed to reduce false negatives in near-OOD settings.

Component	Construction	Intended regime
ENS	MLLM describes negative images richly	Far-OOD
VSNL	MLLM generates negative labels for visually similar ID subset	Near-OOD
Negative images	Mined from test set using $S_{nl}$	Shared input

ENS is produced by prompting an MLLM to describe each negative image richly rather than assigning a single label, thereby generating sentence-level negative text that tightly characterizes OOD semantics (Wenjie et al., 4 Sep 2025). The formal expression is

$$\mathcal{Y}^-_{ens} = \mathcal{G}_{ens}(\mathcal{Y}, \mathcal{X}_{neg}, f_{mllm}, M),$$

where $\mathcal{Y}$ is the ID label set, $S_{nl}(x)$0 is the MLLM, and $S_{nl}(x)$1 is the mined negative-image set (Wenjie et al., 4 Sep 2025). Random or repeated sampling is used to ensure a consistent set size $S_{nl}(x)$2 (Wenjie et al., 4 Sep 2025).

VSNL is motivated by the observation that ENS can misclassify in near-OOD settings because visually similar OOD samples may also fit rich negative descriptions (Wenjie et al., 4 Sep 2025). ANTS therefore first mines the subset of ID classes visually similar to the negative images. For each ID class $S_{nl}(x)$3, it computes

$S_{nl}(x)$4

and then selects a top-$S_{nl}(x)$5 proportion of ID classes with highest $S_{nl}(x)$6 to form $S_{nl}(x)$7; the detailed summary gives $S_{nl}(x)$8 as an example (Wenjie et al., 4 Sep 2025). The MLLM is then prompted with these ID classes and the mined negative images to generate negative labels resembling only this subset: $S_{nl}(x)$9

The distinction between ENS and VSNL is central. ENS emphasizes expressive coverage of novel semantics; VSNL emphasizes local discrimination against a visually confusable subset of ID classes. This division is one of the framework’s main mechanisms for separating far-OOD and near-OOD behavior (Wenjie et al., 4 Sep 2025).

4. Scoring and adaptive balancing

ANTS computes similarity-based scores with the CLIP encoder for ID texts and for both negative textual spaces, then combines the resulting scores with an adaptive weighted rule (Wenjie et al., 4 Sep 2025). The final score is

$$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$0

where $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$1 is computed adaptively (Wenjie et al., 4 Sep 2025).

The adaptive weight is determined from the mean scores of ENS and VSNL over the mined negatives. The detailed formulation is

$$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$2

with

$$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$3

The interpretation given is explicit: if ENS performs better on the mined negatives, $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$4 moves closer to $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$5, favoring ENS for far-OOD; if VSNL performs better, $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$6 moves closer to $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$7, favoring VSNL for near-OOD (Wenjie et al., 4 Sep 2025).

This adaptive balancing is intended to eliminate the need for prior knowledge about whether the current task is near-OOD or far-OOD. A common misunderstanding is that the method must know the OOD regime in advance. The stated design objective is the opposite: it balances the two negative textual spaces automatically and data-drively, “without relying on task-specific prior knowledge” (Wenjie et al., 4 Sep 2025).

5. Empirical performance and comparison with adjacent OOD methods

On ImageNet-1K, ANTS “significantly reduces the FPR95 by 4.2% over the best baseline,” and the abstract describes this as establishing “a new state-of-the-art” (Wenjie et al., 4 Sep 2025). On OpenOOD, the reported results include $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$8 for Far-OOD and $$\mathcal{X}_{neg} = \{ x \mid S_{nl}(x) < \gamma, x \in \mathcal{X}_{test} \}.$$9 for Near-OOD (Wenjie et al., 4 Sep 2025). The framework is further described as robust to ID noise, domain shift, model choice, and prompt variation, and ablations confirm the benefit of both negative image mining and adaptive VSNL construction (Wenjie et al., 4 Sep 2025).

ANTS belongs to a broader family of test-time adaptive OOD detectors based on negative guidance, but its mechanism differs from prior adaptive proxy methods. AdaNeg, for example, addresses semantic misalignment in static negative labels by constructing adaptive negative proxies from a feature memory bank during testing, using task-adaptive proxies that average stored features and sample-adaptive proxies that weight features by similarity to each test sample (Zhang et al., 2024). AdaNeg is also training-free and annotation-free, and on the large-scale ImageNet benchmark it reports a $\gamma$0 increase in AUROC and a $\gamma$1 reduction in FPR95 relative to NegLabel (Zhang et al., 2024). The contrast is structural: AdaNeg uses memory-backed image features as adaptive proxies, whereas ANTS shapes an adaptive textual space by prompting an MLLM with mined negative images (Wenjie et al., 4 Sep 2025).

This comparison helps clarify what is specific about ANTS. It does not merely adapt visual prototypes, and it does not rely on a fixed negative vocabulary. Its main innovation is to use MLLM-generated negative text to approximate the encountered OOD semantics while separately controlling near-OOD failure modes through VSNL (Wenjie et al., 4 Sep 2025).

6. Conceptual placement and terminological scope

ANTS is closely related to earlier work that uses negative textual semantics as a discriminative signal. In universal multi-source domain adaptation, APNE-CLIP introduces “negative textual semantics” through adaptive prompt learning, a memory bank of pairwise textual distances, and energy-based uncertainty modeling to improve image-text alignment and unknown-sample detection (Yang et al., 2024). That work treats negative textual semantics as representations of non-membership relationships between classes, whereas ANTS uses MLLMs to generate negative text from mined negative images and visually similar ID subsets (Wenjie et al., 4 Sep 2025, Yang et al., 2024). This suggests a broader methodological pattern in which textual non-membership cues are increasingly used to sharpen boundaries under distribution shift.

The phrase “negative textual space” also appears in a different sense outside OOD detection. In text-to-image generation, “Orthogonal Negative Guidance in Attention Feature Space” defines the negative textual space as directions in feature space corresponding to undesired concepts that are orthogonal to the intended prompt, and suppresses those directions by orthogonalizing negative-prompt attention features with respect to positive-prompt features (Ko et al., 28 May 2026). That usage concerns concept suppression in MM-DiT-based text-to-image transformers rather than OOD detection. A separate, unrelated paper reuses the acronym ANTS for “Adaptive Nucleus Truncation Sampling,” a long-form reasoning sampler based on standardized logit-space neighborhoods, entropy-conditioned truncation, and a no-truncation fallback arm (Dia, 12 Jun 2026). These parallel usages make the full expansion—Adaptive Negative Textual Space—important for disambiguation.

Within OOD detection, the defining contribution of ANTS is therefore specific: it shapes an adaptive negative textual space by MLLM at inference time, using mined negative images to generate expressive negative sentences and visually similar negative labels, and then balances them with an adaptive score for near- and far-OOD settings (Wenjie et al., 4 Sep 2025). The framework’s training-free and zero-shot character, together with its reported ImageNet and OpenOOD results, position it as a test-time textual adaptation method rather than a finetuning-based or memory-only proxy method (Wenjie et al., 4 Sep 2025).